We built an AI App in Sigma to automate SEO blog content generation. Here’s how we did it

A client approached us with a big idea.

Client: Instead of manually writing our SEO content, can you automate the entire process?

Me: Yes!! (and we did)

This blog walks you through how I built a Sigma AI App that uses Snowflake LLMs to automate marketing workflows.

Why Create Blog Content?

Did you Google it? That’s the first thing that comes to mind for researching something you are unfamiliar with. People search for problems, comparisons, and “how-to” scenarios long before they become product consumers. When your website has clear answers to these questions, it appears more frequently during those searches. Marketers prioritize blogs because they produce measurable returns. For businesses, content tends to deliver value in a few repeatable ways:

• Demand capture: Posts appear when prospects search for answers, supporting lead flow without relying entirely on ads.
• Sales support: Strong articles reduce repetitive “explain it again” work by giving sales and customer success something credible to point to.
• Trust building: Buyers can evaluate your thinking before they ever talk to you, which can shorten cycles and make conversations more efficient.
• Content reuse: One article can be repurposed into social posts, talk tracks, and sales collateral without rewriting from scratch.

Manual blog writing feels like running a relay race with way too many handoffs. Each post requires selecting a topic, conducting keyword research, coordinating with SMEs, briefing, drafting, editing, formatting, and publishing. While no single step in the process is inherently difficult, each one has a different owner, timeline, and definition of done. This is the gap an automated marketing workflow is meant to close: creating tighter continuity from performance signals to content decisions, and fewer dropped balls between “we should write this” and “it shipped.”

LLM Configuration

First, I had to think through how to structure the workflow. I decided on an assembly-line approach with 25+ stages. Each stage has a single responsibility, a defined output, and directly supports a step in the sequence.

Those stages fall into a few families:

• Strategic agents that handle SEO planning and topic ideation. I used claude-3-7 sonnet to generate a prioritized list of SEO keywords.
• Drafting agents that build the outline, assemble research notes, create an outline, and then produce full sections. To accomplish this, I used claude-4-sonnet with a temperature range of 0.3 to 0.65.
• Quality control agents that run checks in parallel, consolidate findings, and prepare a clean set of edits to apply. To do this, I used llama3.3-70b.
• Refinement and editorial agents that handle tone, product callouts, creative alternatives, and final assembly. This was achieved using openai-gpt-4.1 with a temperature of 0.35. 

This architecture keeps early decisions from being overwritten in later stages and makes review and troubleshooting easier. Each stage is configured as an LLM agent with elements that can be tuned independently:

• The model is based on reasoning depth, speed, or long-form coherence. Different tasks behave better with different models, so each agent is configured with a model choice and an expected output.
• The prompt structure uses either a two-prompt split or a self-contained prompt. Prompt structure matters because models do not behave uniformly. When predictability matters, lean on the single prompt pattern. This also keeps the workflow easy to maintain, as self-contained prompts can be revised without inadvertent downstream impact.
• The temperature range is set to match the task type, like verification, structure, and ideation. The stages that must behave predictably are run at lower temperatures, while ideation and creative stages allow more variation.

Sigma Integration

Once the LLM agents are in place, the next step is running them directly from Sigma. That’s where the workflow becomes usable for marketing teams. The design goal is straightforward: give users a structured place to shape an input, add context, adjust tone, and run subsequent stages, without turning the workbook into a giant prompt engineering exercise.

Sigma invokes a Snowflake function that returns VARIANT

CallVariant is Sigma’s way to call a warehouse function when the return type is VARIANT (a semi-structured, JSON-like result).

Here’s an example of the CallVariant function used in the Blog Automation AI App:

Sigma sends the model request as a message payload

SNOWFLAKE.CORTEX.COMPLETE is Snowflake’s Cortex function for calling a supported LLM and returning its response to your prompt.

SNOWFLAKE.CORTEX.COMPLETE is called with:

• a model identifier, and
• an array of messages that follow a chat-style structure, where each message has a role and content.

This is how the workflow supplies instructions and input text to the model in a consistent format.

Prompts are assembled in Sigma before the call

Rather than hardcoding a one-off prompt each time, Sigma builds the final prompt from a template plus the current content being processed. The template includes placeholders, and Sigma replaces them with the current input, allowing the same template to be reused across many runs.

Optional parameters control output behavior

The call includes runtime parameters, such as temperature, to influence how strict or creative the completion should be for that step.

Sigma extracts and displays the completion

Cortex returns the result in a structured format, and the generated output is included in choices. Sigma then selects the relevant part of the payload and converts it to plain text so it can be displayed in a cell or text element, saved as an intermediate artifact, or passed to the next step of the workflow.

How to create an automated blog post in Sigma

Watch the video below as I walk through our Sigma AI App to create SEO blog content!

The final prompt

The shift to an automated marketing workflow turns one-off writing projects into a repeatable sequence of decisions, drafts, checks, and approvals that the team can run, review, and improve. In our case, that operational change produced a measurable outcome: blog production increased 1500% on the same budget that previously supported a fully manual writing process. 

If you’re evaluating a similar approach, start by identifying the parts of your content process that cause the most delays, such as inconsistent briefs or scattered feedback. Pick a workflow that repeats often, build the planning and drafting path, then add quality control and editorial stages once the team trusts the output and the review flow. If you want to take the next step, run a short pilot with one content type, track cycle time and rework rate, and compare output volume and quality against your manual baseline. From there, expand the workflow stage by stage so it grows with the team rather than forcing a single high-risk switch.

Contact Us

If you would like to talk to someone at Maverick Data about maximizing your usage of the Sigma platform, please email us at spencer@maverickdata.io for more information!